What is GPT doing…
By Stephen Wolfram
This is a short book which combines an description of how
GPT, the revolutionary new AI technology shaking the world, works with an advertisement
for some Wolfram products designed to augment GPT. His description is pretty good but I couldn’t
really follow his explanation of the crucial “transformer” component, the
attention blocks. ChatGPT was able to
give me a better account.
Wolfram pointed out some key failings of GPT. It can’t really do deductive chains of any
substance. For example, ask it to
compute 3^73 and it comes up with an absurdly wrong answer, off by a factor of
roughly 10^34. I tried giving it some
hints, including using logarithms.
Again, it confected a plausible looking deductive chain and an answer
only off by a factor of ten or so.
Wolfram then argues that supplementing GPT with some
deductive system like Wolfram language would have more capability. Of course lots of people are working on such
things.
Comments
Post a Comment